Research On Image Feature Extraction Technique Based On Scale Space Theory

With the application of image feature extraction technology in the field of computer vision is more and more important,the requirements of image features which are more robust and better reflect the image content attributes are significantly strict.Compared with the global feature,the local image features are more robust to the geometric and photometry transformations(such as rotation,scale,affine and illumination transformation,etc.)of the image.In addition,the local feature has a very high repeat rate,and they are not susceptible to the influence of target occlusion.Therefore,the research on image feature extraction technology is paid more and more attention.In the image matching application,in order to improve the comprehensive performance(scale invariance,affine invariance and real-time)of image features,this paper makes a detailed study and in-depth analysis of image local feature extraction technology.Firstly,the paper makes an in-depth study on scale space theory,and the basic principle of image feature detection algorithm based on scale space is studied in detail.Including SIFT detector based on Difference of Gaussian(Do G),Harris-Laplace scale invariant detector based on Gaussian scale space and its extended algorithms,Harris-Affine and Hessian-Affine affine invariant feature detectors which can be applied to affine transformation.The application of the scale space in the feature detection ensures that the detected features are scale invariant and have a better stability.Besides,we compared the performance of several feature detection algorithms by carrying out evaluation experiments.Based on the experimental results,we choose the Hessian-Affine detector with the best robustness as the preprocessing algorithm for follow-up feature description.Secondly,from the perspective of image matching,the paper studies the image feature description algorithms in-depth,including MROGH,FRDOH,and LIOP algorithms.According the research results of the principle of the above algorithms,we can improve the robustness and distinctiveness of the feature descriptor by introducing Multi-Support Region idea or dual gradient histogram.However,introducing the Multi-Support Region to construct the descriptors re-sulting in the increase of the computation time in the construction process.Finally,we propose a novel approach(named DGOH)based on dual gradient orientation histogram.We compared the performance of our algorithm with other algorithms by carrying out the evaluation experiments.We can obtain the conclusion that the robustness of DGOH descriptor is better than FRDOH,besides,DGOH has the best real-time performance.